Use PepMLM Online
Commercially Available PepMLM No-Code Web Server
PepMLM: Target Sequence-Conditioned Generation of Peptide Binders with a Protein Language Model
PepMLM is a masked language model that generates de novo linear peptide binders conditioned only on a target protein's amino acid sequence. It was developed by Tianlai Chen, Pranam Chatterjee and colleagues at Duke University, Cornell University and McMaster University, and it is described in an arXiv preprint (2023, revised 2024). PepMLM needs no target structure, which makes it useful for proteins that lack stable folds or accessible binding pockets.
Many disease-relevant proteins are hard to drug. Intrinsically disordered or flexible targets may have no well-defined pocket, and induced-proximity approaches such as PROTACs and molecular glues have their own limits. Structure-based design tools also need a usable structure to begin with. PepMLM tackles this by treating binder design as a sequence-completion problem.
Fact | Detail |
|---|---|
Task | Generate linear peptide binders for a target protein sequence |
Base model | ESM-2 (650M parameters), fully fine-tuned |
Training data | About 10,203 peptide-protein entries from PepNN and Propedia |
Limits in training data | Binders up to 50 residues and targets up to 500 residues |
Structure required | None |
How PepMLM Works: Span Masking on ESM-2
Concatenate sequences: the binder peptide is placed at the C-terminus of the target protein sequence.
Mask the whole binder: a span-masking strategy hides the entire peptide region during fine-tuning.
Learn to reconstruct: ESM-2 is trained to fill in the binder given the target.
Sample peptides: at generation time, top-k sampling with k equal to 3 produces candidate binders, and pseudo-perplexity over the binder region is used to evaluate them.
PepMLM Results: In Silico Benchmarks and Experimental Validation
Comparison with RFdiffusion
In co-folding tests with AlphaFold-Multimer, PepMLM reached an in silico hit rate above 38%, while RFdiffusion stayed below 30% on structured targets. Perplexity correlated negatively with AlphaFold-Multimer confidence (Pearson -0.414 with ipTM and -0.490 with pLDDT).
Target | Assay and outcome |
|---|---|
NCAM1 | ELISA: the top PepMLM peptide bound significantly more strongly than all four RFdiffusion peptides |
MSH3 | 11 peptides tested in HEK293T cells fused to an E3 ligase domain; one induced robust degradation |
Huntingtin (Q43) | 5 peptides tested in patient fibroblasts; 3 showed robust degradation |
Viral phosphoproteins (NiV, HeV, HMPV) | Overall hit rate of about 40%, with several designs reducing the P protein by more than 50% |
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with PepMLM on Tamarind Bio
Binders for hard targets: propose peptides for proteins without a known structure or pocket.
Targeted protein degradation: the authors fused designed peptides to E3 ligase domains to degrade proteins of interest.
Disease-protein targeting: the paper reports peptides against huntingtin and the DNA repair protein MSH3.
Antiviral research: peptides were tested against phosphoproteins from Nipah, Hendra and human metapneumovirus.
How to Use PepMLM on Tamarind Bio
Log in and open the tool: sign in at tamarind.bio and open PepMLM.
Enter the target sequence: paste the amino acid sequence of the protein to bind, up to 500 residues as in the training data.
Set the peptide length: choose the binder length, if exposed. The paper's experimental peptides were roughly 10 to 19 residues.
Set the number of peptides: choose how many candidates to sample, if exposed.
Run the job: submit the run and wait for generation to finish.
Download outputs: retrieve the peptide sequences and their perplexity scores.
Interpret results: lower perplexity indicates sequences the model finds more binder-like, and the paper pairs this with AlphaFold-Multimer co-folding.
Use downstream: co-fold the best candidates and test them experimentally.
Things to Keep in Mind
The peptides in the paper were not experimentally optimized, for example by cyclization or stapling.
Whether huntingtin binders discriminate the mutant Q43 variant from Q17 is not yet established.
The model is not conditioned on post-translational modifications or variants.
Top-k sampling is random, so hit rates can vary between runs, and the ELISA tested only four peptides per method.
The authors note the released PepMLM-650M model on Hugging Face is for academic use only.
Source: Chen T, Dumas M, Watson R, Vincoff S, Peng C, Zhao L, et al., and Chatterjee P. PepMLM: Target Sequence-Conditioned Generation of Peptide Binders via Masked Language Modeling. arXiv:2310.03842 (https://arxiv.org/abs/2310.03842). Code: https://github.com/programmablebio/pepmlm.